If you had access to a robot foundation model, how would you play with it? 🤖 In @GRASPlab , we were fortunate to be early testers of @physical_int Pi0. We found it hard to rigorously benchmark all feasible tasks for a generalist policy. Instead of trying every possible task,

Robot using full size espresso machine? Looks very exciting! I will take a closer look in the technology here

Check the blog for more information https://t.co/4jO3SxBdbl

@Ken_Goldberg 6) @GRASPlab will post the recording on YouTube soon, worth your watching! I really appreciate to study at Penn, where I can actively learn from outstanding people from all over the world. Looking forward to seeing next great insights

@Ken_Goldberg 5) Through a series of concrete examples—including robot table tennis, tactile manipulation, quadruped locomotion, and dynamic motor skill learning on anthropomorphic arms—Peters illustrates how inductive biases enable efficient policy search, reinforcement learning, and

4) It’s very aligned with what @Ken_Goldberg ‘s hot take on GOFE in last week’s talk: solid engineering & science still matter. You need system engineering to drive robot learning.

3) The takeaway is simple yet inspiring : Structure × Learning is the most viable path for robotics. We need engineers to inject their inductive biases to deploy models. Unlearn bad behavior is more important than learning all behaviors.

2) His key point: Inductive biases in robotics aren’t formulas, they are choices about what the robot should learn and what it shouldn’t. Examples: • keep policies close to data distribution • smooth + low-gain control • low-stiffness hardware that adapt well to different

1) The quest for intelligent robots capable of learning complex behaviors from limited data hinges critically on the design and integration of inductive biases—structured assumptions that guide learning and generalization. In this talk, Jan Peters explores the foundational role

We are very happy to welcome @Jan_R_Peters at @GRASPlab , UPenn and give a talk on “Inductive Biases for Robot Learning.” After talk, students like me were fortunate to have lunch together. To be honest, at first, I expected heavy math/physics priors, but his message was

Good day starts with good book, which gives you more visionary insights

High quality product review. Love it

Good way to use video-generation world model in robotics

In my opinion, VLAs research is extremely empirical compared to many other directions. Simulations like LIBERO are no longer statistically meaningful to VLAs, as we can overfit to ~99% easily now. Urgent priorities: 1) Create new sim benchmarks 2) Show real-world experiments

Train GPT to predict the next token and latent. We prove this learns beliefs (sufficient statistic for optimal control) and a world model over beliefs. Results in a powerful GPT, with no changes in architecture or inference. Led by @jayden_teoh_ , applying to PhDs this fall.

Just all in robot learning, where you feel passionate and excited to build anything cool, stay tuned on this “underscore” kid from Urbana!

Generalist Policy for Humanoid!

I’m personally much more excited by mobile, bimanual home robots that can quietly handle chores for a family. But it’s hard to ignore that “humanity-as-a-service”, aka, the attention economy, reaches far more people than the hard work to "make robots really work", Many technical

really cool large datasets maybe the 1st principle data for humanoid foundation policies pre-training, I stay tuned on next crazy work based on it

Dance with the robots!🥰 not fight with them ...

Hi New York, Hi Columbia! Very insightful conversation with Prof. @YunzhuLiYZ and @YXWangBot , @binghao_huang , excited to see next great work from RoboPIL Lab!